arXiv

“SiMDex uses only ~1.49M mined samples (<5% of the pool) yet improves the overall success rate from 47.7% to 61.1%”

“The tokenization compresses robots to 28-101 tokens, which is '27-110× more compact than MJCF text tokens'”

“Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design”

“The paper introduces ATACOM-DC, a method that significantly speeds up the reinforcement learning (RL) process for robots operating under safety constraints.”

“An equivalent and more practical solution consists in setting the corresponding diagonal entries to the upper bound value... critical for anyone implementing this in a modern RL stack like Isaac Gym.”

“Participants: Paolo Magliano, R. Camoriano, et al. (arXiv Physical AI)”

“Our results show that in all tasks Directional Constraints consistently lead to faster learning and overall improved final performance”

“S2-VLA introduces a State-Space Guided Adaptive Attention (SSGAA) mechanism that dynamically shifts focus between spatial perception, task planning, and execution consistency depending on the current phase of the task”

“S2-VLA (2B parameters) outperforms GR00T N1 (2B parameters) on the LIBERO benchmark average (98.2% vs 93.9%)”

“adding this prober improved position prediction accuracy by roughly 3.9x and attitude accuracy by 8.5x compared to an unconstrained neural network decoder (Section VI-B, Table III)”

“SkyJEPA: Learning Long-Horizon World Models for Zero-Shot Sim-to-Real Control of Quadrotors”

“The core achievement of this paper is a framework that recycles a robot's past successful actions to teach it how to handle new objects it has never seen before.”

“By fine-tuning Vision-Language-Action (VLA) policies on this augmented data, they achieved a 16.5% relative to the state-of-the-art baseline on novel objects.”

“arXiv: 2605.13119”